Handcrafted vs. Deep Radiomics vs. Fusion vs. Deep Learning: A Comprehensive Review of Machine Learning -Based Cancer Outcome Prediction in PET and SPECT Imaging

Deep radiomics features outperform traditional handcrafted features: Deep radiomics features (DRF) achieved the highest mean accuracy at 0.862 across cancer outcome prediction studies, suggesting that automatically learned features from imaging data may capture clinically relevant patterns better than manually designed ones.

Fusion approaches show promise but lack standardization: Hybrid fusion models yielded the highest AUC scores (0.861), but only 48% of the 226 studies reviewed adhered to standardized feature extraction guidelines (IBSI standards), raising concerns about reproducibility and clinical reliability.

Critical data quality issues threaten real-world deployment: Nearly 60% of studies inadequately addressed class imbalance, 29% had missing data problems, and 19% showed low population diversity—these gaps could lead to AI systems that perform poorly for underrepresented patient groups or fail in clinical practice.

PET imaging substantially outperforms SPECT for cancer prediction: 95% of studies used PET imaging and generally achieved better results than SPECT-based approaches, likely due to superior spatial resolution and sensitivity, which has implications for which imaging technologies should be prioritized for AI-assisted cancer care.

Explainability remains underdeveloped despite clinical needs: The review emphasizes that interpretable AI is essential for clinical integration, yet the abstract indicates widespread gaps in this area—meaning many current ML models cannot adequately explain their predictions to doctors and patients, hindering trustworthy adoption.